Originally published at terminalblog.com
Related articles
- Best Coding Agents 2026 — Decision Guide
- Coding Agent Observability: How to Log, Monitor, and Debug AI-Generated Code
OpenAI Codex released version 0.146.0 on July 29, and it is one of the biggest updates to the coding agent in months. The release adds named sessions, the ability to pin important threads, side conversations you can switch between, a plugin system for agent extensions, and the ability to fork any thread into a new branch.
If you use Codex regularly and have ever lost track of which terminal window was handling which task, or wished you could try a different approach without losing your current progress, this release addresses both problems.
Named sessions: give each task a label
Before v0.146.0, every Codex session was anonymous. You opened a terminal, started working, and if you opened a second terminal for a different task, you had no easy way to tell them apart. Codex now lets you name sessions using /new or /clear.
When you type /new in a Codex terminal, you get prompted for a name. Call it something descriptive — "refactor-auth-module" or "debug-payment-flow" — and that name appears in the session header and in session lists. The /clear command does the same thing: it starts a fresh session and lets you name it immediately.
For beginners, think of this like naming browser tabs. When you have five Chrome tabs open, you rely on the title to know which is which. Named sessions do the same thing for your coding agent sessions.
You can also pin threads. Pinned threads stay at the top of your session list so the ones you are actively working on never get buried under completed sessions. If you are juggling three or four tasks across a workday, pinning the active ones keeps them one keystroke away.
Side conversations without losing context
The new side conversations feature lets you branch off from an existing thread to ask a follow-up question or explore an alternative approach, without disrupting the main conversation. You switch between the main thread and its side conversations using the thread UI.
This is useful when an agent is working on a task and you want to ask "what if I used a different database schema?" without derailing the current work. The side conversation lives alongside the main thread. You can go back to the original conversation at any point.
The practical effect is that you stop losing context when you want to explore alternatives. Before this feature, developers often opened a new terminal session entirely, losing the conversation history with the agent. Now the branch stays connected to the original.
Thread forking: try a different approach safely
Thread forking goes further than side conversations. When you fork a thread, you create a complete copy of the conversation history up to that point, and the fork starts fresh from there. You can then give the agent new instructions without affecting the original thread.
The forked thread has paginated history so you can scroll back through the conversation even when it gets long. Temporary forks — ones you create just to try something quickly — do not clutter your thread listing. They exist only while you are working in them.
Think of forking like creating a branch in git. You are taking the current state, making a copy, and working from there. If the forked approach does not work out, you close it and go back to the original. If it does work, you have a clean record of the alternative path.
For beginners, the difference between side conversations and forking is straightforward. A side conversation is a quick question attached to an existing thread. A fork is a full copy of the thread that becomes its own independent workspace. Use side conversations for "quick, what about...?" questions and forks for "let me try a completely different approach."
Agent Plugins: extend Codex with third-party tools
The other major addition in v0.146.0 is Agent Plugins support. Codex can now load plugin manifests — configuration files that describe tools, resources, and capabilities a plugin provides. Once a plugin is installed, the agent can use its tools alongside the built-in ones.
The release adds support for workspace plugin publishing, which means a team can create a plugin and distribute it to every developer on the team through a shared workspace. Additional plugin marketplaces for Amazon Bedrock and Claude Code are also supported, so you are not limited to a single source for plugins.
For someone new to coding agents, a plugin is like an app for your agent. Just as you install a browser extension to add functionality to Chrome, you install a Codex plugin to give your agent new capabilities — access to a deployment service, integration with a project management tool, or specialized code analysis features.
Codex also now discovers executor-provided skills. If the underlying code executor has additional capabilities, Codex automatically detects them and makes them available to the agent. This means plugins do not always require explicit configuration — some just work once they are installed.
Remote Code Mode over WebSocket
A quieter but important change: Codex can now connect its app server to remote Code Mode hosts over WebSocket. If you run a remote execution environment — a dev server, a cloud VM, or a containerized workspace — Codex can send commands to it and receive results without requiring a local installation.
This matters for teams that use remote development environments. Instead of running the full agent locally and having it execute code remotely through SSH, the agent communicates directly with the remote host. It is faster and more reliable.
Web search for custom providers
Codex v0.146.0 also adds standalone web search support for compatible custom model providers. If you are running Codex with a non-OpenAI model through a compatible provider, the agent can now search the web for up-to-date information without relying on the default provider's web search implementation.
This is relevant for teams that use open-source models or self-hosted inference. Before this change, web search only worked with OpenAI's default providers.
Bug fixes and quality improvements
The release includes significant stability work. Proxy settings are now honored consistently across authentication, plugin downloads, MCP authorization, remote execution, WebSockets, redirects, and LM Studio connections. Previously, some of these paths ignored proxy configuration entirely, which broke workflows for users behind corporate firewalls.
MCP connections are now kept current when authentication or configuration changes. If you update your API key or change a server configuration, Codex reconnects the affected servers without restarting the ones that are still healthy.
Terminal rendering improved across the board. Nonblocking interrupts mean you can press Ctrl+C without the UI freezing. Narrow layouts render correctly. Keyboard handling on Windows is more reliable, including navigation keys that were previously broken. Sandboxed process trees are now terminated reliably on all platforms.
The skill catalog retention improved too. When context gets tight, Codex now preserves more available skills and warns you when it has to truncate the list instead of silently dropping them.
How to update
Run your normal Codex update command. If you installed Codex through npm:
npm update -g @openai/codex
If you installed through a different method, check the Codex GitHub releases page for v0.146.0. The alpha track already has v0.147.0-alpha.1 if you want to preview what is coming next, but for production use, stick with v0.146.0.
What this means for coding agent users
This release moves Codex closer to being a full development environment rather than a terminal tool. Named sessions, pinned threads, side conversations, and forking are features you expect from an IDE, not a command-line agent. Adding Agent Plugins turns Codex into a platform rather than a standalone tool.
For beginners, the most immediately useful change is named sessions. Give each task a name, pin the ones you are actively working on, and you will never lose track of what is happening in which terminal window again. The rest — plugins, remote Code Mode, web search for custom providers — becomes relevant as your workflow gets more complex.
Top comments (0)